2.092/2.

093 Finite Element Analysis of Solids & Fluids I

Fall 09

Lecture 21 - Solution of the Generalized Eigenvalue Problem

Prof. K. J. Bathe

MIT OpenCourseWare

Reading assignment: Chapters 10 and 11

+ KU = R
MU

(1)

Aside: M could have zero masses. Then we use Gauss elimination on K to remove zero-mass DOFs, but
we denote the nal matrix still as K. Then, in free vibrations:
+ KU = 0
MU

(2)

T MU
> 0, U
T KU
> 0 for any
where now M and K are assumed to be positive denite matrices, i.e. U
=
U
0. Then, we obtain the eigenvalue problem
K = M

Ki = i M i

(A)

where 0 < 1 2 . . . n .

Recall:
Ti M j = ij
Ti Kj = i2 ij = i ij

The Case of Multiple Eigenvalues

Assume 1 = 2 < 3 , i.e. 1 has a multiplicity of 2 (m = 2), 1 and 2 are two eigenvectors for 1 and 2 ,
and 1 =
2 . Then, we have
K1 = 1 M 1
K2 = 1 M 2

(: any constant)
(: any constant)

(3)
(4)

Hence,
K (1 + 2 ) = 1 M (1 + 2 )

(5)

is also an eigenvector corresponding to 1 ! We can change the length of the

Eq. (5) shows 1 + 2 =
eigenvector so that for some ,

T
M
=1

i + i2 xi = ri , having set the mass m to 1 since Ti M j = ij .

Recall we want x
If the eigenvalues for the system (A) are distinct, the eigenvectors are unique. Here, we have a two dimensional
eigenspace (1 = 2 ). Any two M -orthogonal vectors in this space are eigenvectors and could be used as
mode shapes.

Gram-Schmidt (see textbook)

Orthogonalization is used to obtain M -orthogonal vectors. For an eigenvalue of multiplicity m, we have an
eigenspace of dimension m and can always nd m M -orthogonal vectors that are in this eigenspace. We
need orthogonality to decouple Eq. (2). Next, we will discuss some solution techniques.
1

Lecture 21

Solution of the Generalized Eigenvalue Problem

2.092/2.093, Fall 09

Inverse Iteration
Once we have eigenvectors with Ti M j = ij , we could simply use Ti Kj = i ij to obtain i .
Do we need to iterate on K = (M ) to get Ki = i M i ? Since for the general case there are no
explicit formulas available to calculate the roots of p() when the order of p is greater than 4, an iterative
solution method has to be used.

Iteration
Assume 1 > 0. We pick x1 and use for k = 1, 2, . . .
k+1 = M xk
Kx
xk+1 =

(a)

k+1
x
Tk+1 M x
k+1
x

12

Since 1 > 0, K is positive denite and we can solve Eq. (a). We want xk+1 to satisfy the mass orthonor
Tk+1 M x
k+1 = 1. If we assume xT1 M 1 = 0, then
mality relation x
xk+1 1 as k
1 = T1 K1 , T1 M 1 = 1
Proof: Consider
Kxk+1 = M xk

(B)

We see that (B) is equivalent to working with vectors zk+1 and zk .

zk+1 = xk+1

, zk = xk

Substitute into (B):

T Kzk+1 = T M zk

1
zeros

zk+1 = zk
..

.
zeros
n

(C)

Working on (C) is equivalent to working on (B)

Next, iterate with (C). Assume:
z1T =

...

zeros

1
2

..

z2 = z1

zeros

Then we nd
z2T =

1
1

1
2

1
3

...

1
n

After l iterations,
T
zl+1

1
1

1
2

1
3

...

1
n

Lecture 21

Solution of the Generalized Eigenvalue Problem

2.092/2.093, Fall 09

Only the direction of the vector is important.

Assume 1 < 2 . Multiply zl+1 by (1 ) to obtain a new zl+1 :

l l
T
1
1
zl+1 = 1
...

1
n

This zlT+1 converges to [ 1 0

0 . . . 0 ] as l .

1
0


Note that if z1 is orthogonal to 0 , we will never reach the eigenvector corresponding to 1 .
..
.
0

Finally, assume 1 = 2 < 3 . Then we obtain

T
zl+1
=

...

To obtain the 2nd eigenvector for 1 = 2 , choose a starting vector x1 that is M -orthogonal to 1 and
enforce this orthogonality in each iteration. To avoid round-o error, see the textbook.
In practice, the inverse iteration method is hardly used by itself, but rather as an ingredient in a more complex
scheme. The next lecture introduces the widely used subspace iteration method which employs the inverse
iteration method to eciently solve for the rst few lowest frequencies/eigenvalues and modeshapes of large
systems.

MIT OpenCourseWare
http://ocw.mit.edu

2.092 / 2.093 Finite Element Analysis of Solids and Fluids I

Fall 2009

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.